AI in Financial Services Is Moving From Answers to Work Products

Sarah Friar framed Codex less as another interface for querying an AI system than as a move into execution. Her distinction was simple: early use felt like “asking mode,” but repeated use made clearer “how much it goes into this real doing.” In the context of an investor-focused event, that was the operative claim. The productivity story was not only that AI can answer questions faster, but that it can generate work products and participate in workflows where the artifact itself matters.

The event’s on-screen framing identified it as “OpenAI Investor Innovation Day,” shown over a city skyline. OpenAI’s description said it brought together “a curated group of senior leaders for a hands-on experience focused on how AI is transforming financial services.” The remarks stayed close to that practical frame: deal folders, company research, spreadsheets, models, decks, internal GPTs, and enterprise deployment.

A slide made the execution claim concrete. It labeled a step in the process as “Phase 2: Creating An Artifact” and instructed users to ask Codex to build either “a spreadsheet or model” or “a deck,” then add it to the desktop. The same slide included a warning under “Challenge”: “Don’t rush into asking for the artifact.” Artifact generation was presented as one phase in a process, not simply as the first thing to request.

Jasmine Azizi described one practical use case in terms of deal work: connecting a deal folder with its surrounding context and then asking real-time questions against it. The claimed impact was not a generalized improvement in intelligence, but a change to a workflow that depends on accumulated documents and context. In her account, the value comes from making the relevant deal materials available for live questioning when the team needs them.

David Bessel described a related but narrower gain: using ChatGPT’s Excel plugin to speed up the process of “getting smart on a company.” His phrasing points to company research as a recurring task, with Excel as one of the working surfaces where that research becomes usable. The plugin, Bessel said, had sped up that process “very, very meaningfully.”

These examples fit the artifact slide without turning AI into a standalone answer machine. Spreadsheets, models, and decks were presented as artifacts Codex could help produce; deal folders and Excel workflows were presented as places where AI could connect to materials and tools already used in the work. The emphasis was operational and experience-based: make deal context queryable, accelerate company research, and help create the artifacts named in the slide.

James Mackey put the enterprise claim in organizational terms. To “actually change the enterprise,” he said, people need to solve “really hard problems” in order to unlock full organizational transformations. The point was not simply that employees should use AI tools more often. It was that transformation depends on applying the tools to problems substantial enough to affect how the organization works.

Mackey later gave the clearest scale marker in the source: 2,700 employees now have an enterprise ChatGPT license. He described that deployment as “a giant force multiplier for the business.” The mechanism he named was a flywheel: hundreds of GPTs being created by employees and pushed into day-to-day business operations, streamlining what the company is able to do.

The language of a “flywheel” matters because Mackey was describing cumulative internal use rather than a single application. Employees were creating GPTs across the business, and those GPTs were moving into day-to-day operations. His claim was that the accumulation of those uses was streamlining what the company could do.

That connects back to his earlier point about hard problems. In Mackey’s account, broad license access matters insofar as it lets employees apply ChatGPT to meaningful business problems and operational workflows, not merely because adoption has reached a large number.

The source’s strongest through-line was that AI becomes more consequential when it is attached to business context and used inside day-to-day work. The financial-services frame came from OpenAI’s description; the examples were narrower and more concrete: a deal folder with context that can be questioned in real time, an Excel plugin that speeds up getting smart on a company, Codex prompts that produce spreadsheets, models, or decks, and hundreds of GPTs moving into business operations.

The slide’s warning not to rush into the artifact qualified the speed story. Codex was being presented as a way to create work products, but the request for the spreadsheet, model, or deck was framed as part of a phase. Azizi’s example pointed to the need for connected context. Bessel’s example pointed to a familiar analytical tool. Mackey’s example pointed to enterprise access and repeated internal creation of GPTs.

OpenAI’s investor event therefore positioned Codex and enterprise ChatGPT as tools for execution in financial-services work, but the claims in the source were concrete rather than sweeping: query deal context, accelerate company research in Excel, generate spreadsheets and decks, and distribute internal GPTs across day-to-day operations.